When photographer Adri Tan’s family offered to take their
father anywhere in the world to celebrate his 60th birthday, he chose hiking the Alta Via 1 trail (AV1). End to end, the legendary trek in the Dolomites crosses 20km (75 miles), with rocky scrambles,
some stiff ascents and truly astonishing scenery. Because of the strenuous
nature of the hike, a few family members opted to stay home, so Tan’s best
friend accompanied them.
Advertisement
“My dad chose the trail, but I handled all of the logistics,” Tan says. To get to the trailhead in the Dolomites, they flew to Venice, took a bus from the airport to Cortina d’Ampezzo, another bus to Dobbiaco and, the following morning, a third bus to Lago di Braies, where the AV1 begins. Tan booked overnight accomodations in rifugialong the path; this network of mountain huts reaches throughout the Dolomites. The group appreciated leaving the planning to Tan, who speaks Italian and could handle the correspondence with the rifugi easily.
Above from left: Lago di Braies and resting on the trail. Adri Tan (2)The AV1 trail on day 3 of the hike from Rifugio Lavarella to Rifugio Lagazuoi. Adri Tan
The scale of the trip was a major departure for Tan. “I had neither been on a hike with my dad before nor been on an international trip with him outside of visiting family in Malaysia,” says Tan, who was the least experienced hiker of their group. And the trail itself was “grueling. I consider myself to be a fairly active and strong person, but normal cardio and strength training doesn’t prepare you for the endurance required to hike 6 to 11 miles [9km to 18km] a day at a high elevation gain.”
At times, each hiker struggled, Tan says, “except for my best friend, who seemed to have legs of steel.” The group always made it to each night’s rifugio on time though, so their individual challenges didn’t result in any serious delays. “If anything, our more experienced hikers helped us stick to the correct trails and find the ones that had been washed away by rockslides earlier that spring,“ Tan says.
From left: The photographer’s friend on the hike and a cow on the trail. Adri Tan (2)
Advertisement
“I was born in the Year of the Ox, according to the Chinese zodiac, so my family’s nickname for me is Cow or Cow Cow,” Tan says. “We saw a lot of dairy cows grazing along the AV1 trail, and my family got a kick out of calling them my cousins whenever we saw them.”
The photographer’s friend and brother overlooking a body of water at Rifugio Lavarella. Adri Tan
“The Europeans we encountered were very unfamiliar with the existence of Asian Americans,” Tan says. “We hiked during a very busy time of the year in the Dolomites, when mostly every rifugio is booked. The majority of the people we encountered were also foreigners of various nationalities, but we only ever encountered around six other people of color during the trip. It was a very stark contrast coming from Queens, New York, which is the most ethnically diverse place I’ve ever lived. This lack of racial diversity is not an uncommon occurrence in my experiences hiking in the US, but my self-awareness was perhaps amplified by being in another country. Throughout the trip, I became hypervigilant to racial microaggressions, which became very draining. Despite that, hiking the AV1 trail was still an overwhelmingly positive experience for me.”
A view of the mountains through a window at Rifugio Scoiattoli. Adri Tan
“My family didn’t travel often when I was a child, but when we did, we mostly took road trips around the US. My earliest road trip with my family was to Yellowstone as a 2-year-old, where there was photo evidence of me trying to run away while a family photo was being taken,” Tan says.
Another memorable trip was driving from Minnesota to New York City when Tan was 11. “I loved [New York] but can only remember certain moments, mostly of me being tired and hot,” Tan says. “I remember that people were selling turtles on top of plastic water bottles on the street in Chinatown in the blazing sun, and I think I also had a meltdown on Canal Street when my mom refused to buy an expensive New York City zip-up hoodie that I wanted so badly. I didn’t get the hoodie in the end.”
From left: Rifugio Biella and hiking out the next day. Adri Tan (2)
For Tan, hiking in the Dolomites required the same kind of physical and emotional persistence as a long road trip, and the 11-day trail had its highs and lows. “Highs being the most gorgeous views I’ve ever seen, and the lows being hiking up steep trails while hungry and tired. I was pleasantly surprised by how well we all got along, despite the challenging hike,” Tan says.
Above: The photographer’s father, the inspiration and motivation for the trip, at Rifugio Vazzoler. Adri TanAbove from left: An evening scene in the mountains and the photographer’s father on the Alta Via 1 trail. Adri Tan (2)The photographer’s self-portrait shadow. Adri Tan
Adri Tan is a photographer and zine artist in New York. They created a series of zines about their Dolomites hiking experience.
LONDON (Reuters) – More than 50 people protesting against Britain’s decision to ban the Palestine Action group were arrested outside parliament on Saturday, London’s Metropolitan Police said.
Officers made arrests after crowds, waving placards expressing support for the group, gathered in Parliament Square, the force said on X.
In July, British lawmakers banned Palestine Action under anti-terrorism legislation after some of its members broke into a Royal Air Force base and damaged planes in protest against Britain’s support for Israel.
The ban makes it a crime to be a member of the group, carrying a maximum sentence of 14 years in prison.
The co-founder of Palestine Action, Huda Ammori, last week won a bid to bring a legal challenge against the ban.
We retrieved a total of 9342 studies. After removing duplicates, we identified 8018 studies for title and abstract screening. Of these, 35 studies were potentially eligible. Ultimately, we included 23 studies according to the inclusion criteria (Fig. 1) [12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34].
Fig. 1
Characteristics of the included studies
Twenty-three eligible studies with 25 prediction models were included. Ten prediction models were developed on the basis of traditional meta-analysis, whereas the remaining 15 were developed on the basis of individual participant meta-analysis. The characteristics of the included studies are presented in Tables 1 and 2. These studies were published between 2013 and 2023. The predicted outcomes focused on complications of diabetes mellitus (e.g., early or recurrent foot ulceration, early diabetic kidney disease, early diabetic peripheral neuropathy, diabetic retinopathy, and end-stage renal disease), respiratory disease (e.g., bronchopulmonary dysplasia, bacterial rhinosinusitis, radiation pneumonitis, and severe COVID-19), gestational diabetes mellitus, preterm birth, early-onset colorectal cancer, anastomotic leakage, coronary artery disease, 10-year mortality, survival with good neurologic outcome, mild cognitive impairment, seizure, psychosis and the response to placebo in antidepressant trials. Nearly half of the prediction models were developed with data sourced from cohort studies (n = 10), and the remaining models had a mixed study design. A greater proportion of studies focused on multiple ethnicities (n = 13) and assessed the risk of bias for primary studies (n = 14). Three prediction models were developed with sample sizes exceeding one million, and seven prediction models were developed with sample sizes exceeding ten thousand. Twelve prediction models used data from cohort studies to externally validate their predictive performance (Appendix Table 1).
Table 1 Characteristics of the included studies (prediction models based on traditional meta-analysis)
Table 2 Characteristics of the included studies (prediction models on the basis of individual patient data meta-analysis)
Summary of prediction models
Appendix Table 2 presents the main findings of the included prediction models. Five studies developed risk scoring models according to Sullivan’s method. Two studies developed a Rothman-Keller model. The remaining studies constructed logistic regression models (n = 13), Cox proportional hazards models (n = 2) and parametric survival models (n = 1). The number of risk factors included in the prediction model ranged from 3 to 16. Twenty-three prediction models reported AUCs, with a median value of 0.77 (ranging from 0.59 to 0.91), eleven prediction models were calibrated and two models conducted decision curve analysis. The criteria for the exclusion of risk factors from the prediction models included high heterogeneity between studies, no statistically significant pooled results for risk factors, different definitions of risk factors, the exclusion of some risk factors (e.g., p > 0.0001) to develop a parsimonious model, and the use of the Akaike information criterion to remove the least contributing predictors.
Key steps in developing risk prediction models on the basis of meta-analysis
According to the included studies, we summarized several key steps in developing prediction models on the basis of meta-analysis (Fig. 2).
Fig. 2
Key steps in developing prediction models on the basis of meta-analysis
Step I: confirm that the risk prediction model meets clinical needs before being started
The clinical purpose of the risk model, as well as the target population and clinical setting, was defined. If the risk model meets the clinical need, we need to check the research status of prediction models in the field, including data sources for model development, the clinical setting, predictors, external validation, and ease of use. If the existing prediction models do not meet clinical needs, a new prediction model is needed. Before initiating model development, a detailed study protocol should be formulated and prospectively registered. Additionally, assembling a multidisciplinary expert team is essential. This team plays a vital role in iteratively refining the study protocol and provides ongoing methodological and domain-specific expertise throughout the entire research process.
Step II: data collection via traditional meta-analysis or individual patient data meta-analysis
The two main methods of data collection for the development of prediction models are traditional meta-analysis and individual patient data (IPD) meta-analysis. The two methods include the following steps: defining the review question, drafting the protocol, developing a search strategy, selecting studies, extracting data, assessing the risk of bias, synthesizing evidence and performing a meta-analysis, presenting findings and interpreting the results. Step II is pivotal for the development of prediction models on the basis of meta-analysis. The aim is to systematically search for existing predictors and identify those that should be included in the prediction model and obtained their pooled values and 95% CIs. The included studies used various methods to identify predictors including P-values of pooled values based on meta-analysis, stepwise regression, clinical experience and expert consensus. In addition to these methods, for prediction models based on traditional meta-analysis, we recommend using directed acyclic graphs (DAGs) to identify candidate predictors, as DAGs provide a causal framework that guides variable selection and mitigates issues arising from overreliance on mere statistical associations [35]. For prediction model based on IPD meta-analysis, Lasso regression is suggested to select predictors because it selectively highlights important predictors and efficiently removes less critical ones by shrinking their coefficients to zero [36]. Finally, a Delphi questionnaire was used to consult with experienced experts to determine the predictors ultimately included in the prediction model. In this step, it is important to note that the predicted outcomes, predictors and focused population should be defined a priori to reduce heterogeneity between studies.
Step III: risk prediction model development
First, all the predictors included in the prediction model and their pooled values and 95% CIs are obtained from step II. Then, the corresponding β coefficients are calculated on the basis of the pooled data, and a logistic regression model or Cox regression model is constructed. Additionally, if a risk score prediction model is constructed, we can calculate the score for each predictor according to Sullivan’s method or Rothman’s method [37, 38]. Most of the included studies developed a simple risk score model, which is convenient for clinical practice.
Step IV: validation of the performance of the risk prediction model
Model performance can be evaluated using these indexes: discrimination (e.g., C statistics, sensitivity, and specificity), calibration (e.g., calibration-in-the-large, calibration slope, and calibration plot) and clinical utility (e.g., decision curve analysis). In addition, we can use a risk threshold to turn the risk model into a classification rule. For example, according to the optimal cumulative score based on the Youden index, patients were segmented into different risk groups, and Kaplan-Meier curves were used to calculate the cumulative risk of morbidity in different groups. The prediction model is usually evaluated first with internal validation and then with external validation. However, the models developed on the basis of meta-analysis are rarely validated internally, which may be related to the difficulty in obtaining complete individual data during model development. Approximately half of the studies conducted external validation using data from single-center prospective or retrospective cohort studies.
Step V: model presentation andinterpretation
Once the prediction model is developed, researchers need to consider how to present it. Prediction models based on meta-analysis often provide an exact formula for users to validate or use and are simplified in formats such as nomograms, point score systems and websites. To better explain the prediction model, researchers can use examples to demonstrate how to use the model.
Step VI: model reporting
The prediction models usually follow the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD) reporting guidelines [39]. Of the included studies, only five prediction models were reported according to the TRIPOD guidelines [15, 26,27,28, 32]. However, since the models were developed on the basis of meta-analysis, we recommend that the PRISMA statement should also be followed [40].
Congratulations to the Athletes of the Day on Saturday at The World Games 2025 in Chengdu, China. Wushu fighter Zhuoling Lu, who claimed China’s first gold medal at TWG 2025, and Andrea Maya Becerra, two-time medallist and Mexican archer, were both honoured on stage at The World Games Plaza..
Never Give Up – Andrea Maya Becerra
Andrea Maya Becerra of Mexico staged a remarkable comeback to win gold in the women’s compound final at the World Games 2025, edging out Estonia’s Lisell Jaatma 147–146 at Qinglong Lake Park on Saturday.
The 25-year-old, who also competes in the mixed team event, had taken silver the day before with teammate Sebastián García, missing gold by just one ring. In Saturday’s final, the story reversed—she won by a single ring despite a rocky start, dropping two rings in the first round after an unexpected 8 on her third arrow, which drew an audible gasp from the crowd.
Jaatma shot a perfect 10 on all three arrows of the first round, but Becerra fought back, levelling the score at 117–117 midway through the fourth round. Her third arrow of that round put her in the lead for the first time.
“I’ve been nervous the whole time. This was a tough match. Although I was behind early on, I didn’t let myself relax until the very end. I’m happy I did it,” she said, proudly holding her gold medal.
Kick-off for the Home Team – Zhuoling Lu
Tears of joy flowed as Lu Zhuoling captured the first gold medal for host nation China at The World Games 2025—just days after her 22nd birthday.
“This is the best birthday gift I could have dreamed of,” she said after winning gold in the women’s Wushu Taolu Taijiquan–Taijijian on Friday in her international debut.
Lu opened the day with a Taijiquan score of 9.796, building a strong lead heading into the evening’s Taijijian final. Despite placing third in the second event, her morning advantage secured the overall title with a combined score of 19.522.
“Before going on stage, I was a bit nervous,” she admitted. “But I performed as expected—just like in training. I’m grateful for the support and encouragement from the audience.”
The crowd at Hi-Tech Zone Sports Centre Gymnasium responded with thunderous applause, celebrating the first Chinese champion of The World Games 2025.
Germany has announced a suspension of arms exports to Israel due to the worsening humanitarian crisis in Gaza and Israel’s plans to expand military control over the enclave. This move is driven by growing public outcry, as Germany’s come-what-may support for Israel, rooted in its historical guilt over the Nazi Holocaust, is being tested like never before. The suspension of arms deliveries would affect only those that could be used in Gaza.
An ARD-Deutschland TREND survey released on Thursday showed that 66% of Germans want their government to put more pressure on Israel to change its behavior. This is higher than April 2024, when 57% of Germans believed their government should criticise Israel more strongly than before for its actions in Gaza. Despite Germany helping air drop aid to Gaza, 47% of Germans think their government is doing too little for Palestinians there, against 39% who disagree with this. Most strikingly perhaps, only 31% of Germans feel they have a bigger responsibility for Israel due to their history – a core tenet of German foreign policy – while 62% do not.
Germany’s political establishment has cited its approach, known as the “Staatsraison”, as a special responsibility for Israel after the Nazi Holocaust. Merz’s Foreign Minister Johann Wadephul told Die Zeit newspaper that Berlin could not be a “neutral mediator” because they are partisan. However, Merz’s junior coalition partner, the Social Democrats (SPD), had already been more explicit in wanting to put sanctions against Israel on the table.
The deepening divide within Germany has also played out in its media landscape. Der Spiegel magazine accused Israel of violating international humanitarian law and condemned what it said was the German government’s complicity. Meanwhile, Bild, Germany’s largest media group, decried the lack of outrage toward Islamist Hamas whose cross-border assault on Israeli communities triggered the war, pointing to growing anti-Israel sentiment and one-sided protests.
Critics have argued that Germany’s approach has been overly hesitant, weakening the West’s collective ability to apply meaningful pressure for an end to the fighting and restrictions on humanitarian aid to the Israeli-besieged enclave.
(Based on a Reuters report by Riham Alkousaa and Matthias Williams; additional reporting by Lili Bayer in Brussels)
Wanda Woods enrolled because her father advised that typing proficiency would lead to jobs. Sure enough, the federal Environmental Protection Agency hired her as an after-school worker while she was still a junior.
Her supervisor “sat me down and put me on a machine called a word processor,” Ms. Woods, now 67, recalled. “It was big and bulky and used magnetic cards to store information. I thought, ‘I kinda like this.’”
Decades later, she was still liking it. In 2012 — the first year that more than half of Americans over 65 used the internet — she started a computer training business.
Now she is an instructor with Senior Planet in Denver, an AARP-supported effort to help older people learn and stay abreast of technology. Ms. Woods has no plans to retire. Staying involved with tech “keeps me in the know, too,” she said.
Some neuroscientists researching the effects of technology on older adults are inclined to agree. The first cohort of seniors to have contended — not always enthusiastically — with a digital society has reached the age when cognitive impairment becomes more common.
Given decades of alarms about technology’s threats to our brains and well-being — sometimes called “digital dementia” — one might expect to start seeing negative effects.
Thank you for your patience while we verify access. If you are in Reader mode please exit and log into your Times account, or subscribe for all of The Times.
Thank you for your patience while we verify access.
Welcome back to the Abstract! Here are the studies this week that gave me hope, sent me back in time, and dragged me onto the dance-floor.
First, what’s your favorite cockatoo dance move? To be fully informed in your response, you will need to review the latest literature on innovations in avian choreography. Then: salvation for sea stars, a tooth extraction you’ll actually like, ancient vortex planets, and what to expect when you’re an expecting cockroach.
Everybody do the cockatoo
Lubke, Natasha et al. “Dance behaviour in cockatoos: Implications for cognitive processes and welfare.” PLOS One.
If you play your cards right as a scientist, you can spend all day watching cockatoos dance online and IRL. That’s what one team of researchers figured out, according to a new study that identified 17 cockatoo dance moves previously unknown to science.
“Anecdotally, parrots (Psittaciformes) have been reported to show ‘dancing’ behaviour to music in captivity which has been supported by studies on a few individuals,” said researchers led by Natasha Lubke of Charles Sturt University. “However, to date it remains unclear why parrots show dance behavior in response to music in captivity when birds are not courting or in the absence of any potential sexual partner.” Cockatoos, by the way, are a type of parrot.
It’s worth pursuing this mystery in part because parrots are popular pets and zoo attractions that require environmental enrichment for their welfare while in captivity. Listening to music and dancing could provide much-needed stimulation for these smart, social animals.
To that end, the authors watched dozens of videos of cockatoos on YouTube, TikTok, and Instagram, with search terms like “birds dancing Elvis,” “bird dancing to rap music” and “bird dancing to rock music.” They also played music and podcasts to a group of captive birds—two sulphur crested cockatoos (Cacatua galerita), two Major Mitchell cockatoos (Lophochroa leadbeateri) and two galahs (Eolophus roseicapilla)—housed at Wagga Wagga Zoo in Australia.
Illustration of the 10 most common recorded dance movements. Ethogram descriptors based on Keehn et al. [3] and illustrations by Zenna Lugosi. Image: Lubke et al., 2025, PLOS One, CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
The results expanded the existing database of cockatoo dance moves from classics like headbang, foot-lift, and body roll to include new-wave choreography like jump turn, downward walk, and fluff (wherein “feathers are fluffed” in a “fluffing event” according to the study).
All the birds that the team studied onsite at the zoo also danced at least once to audio playback of the song “The Nights” by Avicii. They even danced when music was not playing, bopping around to silence or to tips from the financial podcast “She’s on the Money.”
“Dance behaviour is perhaps a more common behaviour in cockatoos than previously thought,” the team concluded. “Further research is required to determine the motivational basis for this behaviour in captivity.”
It will be interesting to see what forthcoming studies reveal, but my own prediction is that the motivational basis falls under Lady Gaga’s edict to “Just Dance.”
In other news…
Solving the mystery of what’s killing billions of sea stars
Prentice, Melanie et al. “Vibrio pectenicida strain FHCF-3 is a causative agent of sea star wasting disease.” Nature Ecology and Evolution.
Over the past decade, a devastating illness has killed off billions of sea stars in what is the largest marine epidemic on record. Scientists have finally identified the culprit that causes sea star wasting disease (SSWD) as the bacteria Vibrio pectenicida, which is from the same family that causes cholera in humans (Vibrio cholerae).
Sea stars infected with SSWD form lesions and rapidly disintegrate into goo in mass mortality events that have upended ecosystems on the Pacific coast from Alaska to Mexico. The isolation of the agent involved in these grotesque die-offs will hopefully help restore these vital keystone species.
Hakai Institute research scientist Alyssa Gehman checks on an adult sunflower sea star in the US Geological Survey’s Marrowstone Marine Field Station in Washington State. Image: Kristina Blanchflower/Hakai Institute
“This discovery will enable recovery efforts for sea stars and the ecosystems affected by their decline,” said researchers led by Melanie Prentice of the Hakai Institute and the University of British Columbia.
Psst…you have some ancient atmosphere stuck in your teeth
Feng, Dingsu et al. “Mesozoic atmospheric CO2 concentrations reconstructed from dinosaur tooth enamel.” Proceedings of the National Academy of Sciences.
For the first time, scientists have reconstructed atmospheres that existed more than 100 million years ago by studying the teeth of dinosaurs that breathed in this bygone air.
A team analyzed oxygen remnants preserved in the dental enamel of roughly two dozen dinosaur teeth including sauropods (such as Camarasaurus), theropods (including Tyrannosaurus), and the ornithischian Edmontosaurus (go Oilers). This data enabled them to infer carbon dioxide concentrations of around 1,200 parts per million (ppm) and 750 ppm in the Jurassic and Cretaceous periods, respectively.
This is in line with other findings that have found wild swings in CO2 levels during the dinosaur age, likely due to volcanic activity. Earth’s current atmosphere is about 430 ppm, and is rapidly rising due to human-driven greenhouse gas emissions.
“Fossil tooth enamel can thus serve as a robust time capsule for ancient air [oxygen] isotope compositions,” said researchers led by Dingsu Feng of the University of Göttingen. “This novel form of analysis can “provide insights into past atmospheric greenhouse gas content and global primary productivity.”
Vortex planets from the dawn of light
Eriksson, Linn E J et al. “Planets and planetesimals at cosmic dawn: Vortices as planetary nurseries.” Monthly Notices of the Royal Astronomical Society
The first planets ever born in the universe may have formed in vortices around ancient stars more than 13.6 billion years ago. These stars were made of light elements, such as hydrogen and helium, but each new generation forged an itty-bit of heavier elements in their bellies that could potentially provide basic planetary building blocks.
By running simulations of this early epoch, known as cosmic dawn, researchers led by Linn E.J. Eriksson of the American Museum of Natural History found that small rocky worlds, on the scale of Mercury or Mars, could coalesce from dust and pebbles trapped in so-called “vortices,” which are like cosmic eddies that form in disks around newborn stars.
As a consequence, this “suggests that vortices could trigger the formation of the first generation of planets and planetesimals in the universe,” the team said.
Congratulations to everyone who had “ancient vortex planets from cosmic dawn” on their bingo card this week.
Wash it all down with a glass of cockroach milk
Frigard, Ronja et al. “Daily activity rhythms, sleep and pregnancy are fundamentally related in the Pacific beetle mimic cockroach, Diploptera punctata.” Journal of Experimental Biology.
We began with cockatoos and we’ll close with cockroaches. Scientists have been bothering sleepy pregnant cockroaches, according to a new study on the Pacific beetle mimic cockroach, which is one of the few insects that produces milk and gives birth to live young.
“To our knowledge, no study has investigated the direct relationship between sleep and pregnancy in invertebrates, which leaves open the questions: do pregnant individuals follow similar sleep and activity patterns to their non-pregnant counterparts, and how important is sleep for successful pregnancy?” said researchers led by Ronja Frigard of the University of Cincinnati.
Biologists found that pregnant cockroaches need more sleep and those that are sleep-deprived have babies that require longer gestation to develop. Image: Andrew Higley
As it turns out, it’s very important! The team disrupted pregnant cockroaches by shaking their containers four times during their sleeping period for weeks on end. While the well-rested control group averaged 70 days for its gestation period, the sleep-deprived group took over 90 days to deliver their young. In addition, “when chronic sleep disturbance occurs, milk protein levels decline, decreasing nutrients available to the embryos during development,” the team concluded.
For those of us who have been woken up at night by the scuttling of cockroaches, this study is our revenge. Enjoy it while you can, because the smart money is on cockroaches outliving us all.
Machine learning holds promise for optimizing treatment strategies and potentially improving outcomes in respiratory failure but future research and development are necessary to fully realize its potential in clinical practice. | Image Credit: Selvi – stock.adobe.com
Machine learning (ML) holds significant potential for predicting and improving outcomes in acute respiratory failure by leveraging vast amounts of patient data, but its successful integration into clinical practice is contingent upon overcoming challenges related to data quality, system heterogeneity, clinician acceptance, and ensuring health equity, according to a study published in Critical Care.1
The global rise in acute respiratory failure (ARF) has made mechanical ventilation an increasingly critical, but risky, treatment.2 In intensive care units (ICUs), where 35% to 50% of patients require this support, ARF is associated with a high mortality rate of 67.2%. The treatment itself can be dangerous, with lung-related complications from mechanical ventilation contributing to 40% of in-hospital deaths.
Invasive mechanical ventilation is not only expensive, with ICU costs averaging $2300 per day and increasing to over $3900 after 4 days, but it also takes a heavy toll on patients and their families.3 To address this, researchers investigated whether ML could improve the prediction of respiratory failure.1 An expert panel reviewed existing literature and discussed how ML could be applied to better forecast the onset and progression of this condition.
ML and Data Integration in Clinical Settings
Hospitals need to combine data from many sources to manage patient information effectively.1 Integrating large language models (LLMs) could help by incorporating unstructured data like clinical notes, thereby improving predictive capabilities and clinical decision-making.
Both ML and deep learning are being used in modern clinical decision support systems. While ML models have been applied to predict ARF during invasive ventilation, challenges remain with existing deep learning models. Key barriers to implementation include integrating these analytic platforms into electronic health records (EHRs), gaining clinician acceptance, and maintaining the models’ accuracy across different patient populations and clinical practices.
ML’s Clinical Utility for Respiratory Outcomes
During their discussion on critical respiratory outcomes, panelists reached a consensus that predicting the emergence and progression of respiratory failure represented the most actionable objective. The requirement for invasive mechanical ventilation was also identified as a key outcome of interest. A primary priority was to emphasize the significant impact of having early knowledge of a patient’s likelihood of respiratory failure progression. Such advanced warning would provide clinicians with sufficient lead time to collect and interpret targeted diagnostics, thereby enabling potential interventions to avert clinical deterioration.
“There was debate regarding the optimal prediction horizon length, but most agreed that 12 to 24 hours before the onset of respiratory failure would be a useful window to implement preventative strategies,” study authors found.
Effective clinical collaboration and setting appropriate risk thresholds are vital for good decision-making. By only notifying clinicians at significant thresholds, alarm fatigue and over-notification can be avoided. This approach allows for predicting a patient’s response to interventions, which can inform both clinical decisions and trial design. The most valuable strategy, panelists agreed, is to predict and track a patient’s progression through different levels of respiratory support.
Barriers to the Development and Implementation of ML Models
Deep learning models face challenges at both the patient and system levels. Patient-level difficulties arise from the wide variety of underlying pathologies, types of respiratory failure, and treatment approaches, which complicate model development. For ML models, inconsistencies in the source, format, frequency, and quality of data can compromise accuracy. Furthermore, emergency procedures like intubation are often not accurately recorded or time-stamped in EHRs, making it difficult to properly classify events and ensure model precision.
Substantial heterogeneity across health systems, regional practice patterns, and resource availability complicates the development of generalizable ML models. Successful integration and deployment are contingent upon a safe, efficient, and adequate interface between the EHR and the analytics platform.
Physician hesitancy regarding the deployment of AI models in the ICU environment is a significant barrier. This can be mitigated by promoting transparency and avoiding “black box” models, which can foster greater clinician acceptance and collaboration. Consequently, a robust system for continuous monitoring and iterative improvement of model performance is a necessary component of successful implementation.
Strategies for Successful ML Integration and Validation
Successful ML models must not only perform well initially but also maintain their effectiveness across diverse systems and patient groups. To prepare for full clinical deployment, it’s necessary to implement strategies that enhance predictive abilities, detect bugs, and evaluate false positives and negatives. The panel emphasized that a key step in evaluating a model’s effectiveness is to compare its performance in a clinical setting against the existing standard of care. To gain wider acceptance, panelists suggested that future trial designs should prioritize prospective, multicenter studies.
Overcoming Bias and Resource Gaps in AI for Health Care
A key concern regarding technological advancements in health care is their potential to disproportionately impact socioeconomically disadvantaged patients who have limited access to supporting resources and infrastructure. Nevertheless, when ML algorithms are appropriately designed, they possess the potential to advance health equity rather than exacerbate existing disparities.
Further considerations from a health equity perspective involve the risk that health systems with fewer resources may be delayed in adopting these new technologies, thereby widening health disparities. Additionally, available data may be inherently biased. While natural language processing and LLMs can assist in analyzing progress notes, implicit bias can also manifest through subtle linguistic patterns. Consequently, it is imperative to address these health disparities during the development of any ML model to ensure equitable performance and mitigate bias.
“Enhancing predictive capabilities through ML could facilitate a more proactive approach to patient care, potentially improving outcomes. However, many challenges must be addressed to achieve meaningful integration into clinical care,” study authors concluded.
References
Pearce AK, Nemati S, Goligher EC, et al. Can we predict the future of respiratory failure prediction? Crit Care. 2025;29(1). doi:10.1186/s13054-025-05484-7
Demem K, Tesfahun E, Nigussie F, Shibabaw AT, Ayenew T, Messelu MA. Time to death and its predictors among adult patients on mechanical ventilation admitted to intensive care units in West Amhara comprehensive specialized hospitals, Ethiopia: a retrospective follow-up study. BMC Anesthesiol. 2024;24(1). doi:10.1186/s12871-024-02495-9
Getting patients off the ventilator faster: facilitator guide. Agency for Healthcare Research and Quality. February 2017. Accessed August 8, 2025. https://www.ahrq.gov/hai/tools/mvp/modules/vae/overview-off-ventilator-fac-guide.html
Atomic Linux distributions are all the rage at the moment because they offer enhanced security, stability, and simplified update mechanisms.
But what is an atomic distribution?
Let’s just get right to the heart of this particular situation. An atomic Linux distribution is all about how applications are installed and upgraded. Let me paint a picture.
You’re happily using your operating system when an upgrade notification appears. You innocently click to allow the update to happen, but during the process, something goes wrong (which is actually rare with Linux). After the upgrade completes (with you not knowing something went wrong), you reboot, only to find the OS no longer works properly.
What do you do?
If you were using an atomic distribution, the OS wouldn’t allow the upgrade to happen. In other words, it’s an all or none scenario, and if an upgrade cannot happen without issue, it won’t happen at all. Even if the upgrade happens, you can always select a viable snapshot during boot.
What this does is ensure you always have a working instance of your OS.
Yeah, atomic distributions are a smart way to go.
But they aren’t just about upgrades. Another very important aspect of atomic distributions is that the core system is read-only, which means the base files and configurations cannot be directly modified. That’s all about security, and it works very well.
Now, let’s talk about HeliumOS.
What Is HeliumOS?
HeliumOS takes Red Hat‘s CentOS Stream, rebases it with AlmaLinux, and gives it an atomic spin. In that regard, HeliumOS offers:
A version of CentOS that includes the KDE Plasma desktop for a user-friendly, highly customizable experience.
An app store that is comprised of Flatpak packages for up-to-date, sandboxed apps.
A promise of 10 years of support with new features, bug fixes, and security updates.
An atomic desktop OS that doesn’t get in the way.
You’re probably thinking, “But aren’t CentOS and AlmaLinux server OSes?” Why, yes, they are. Does that, in turn, make HeliumOS a server OS?
No, it doesn’t.
In fact, HeliumOS is marketed as an atomic desktop OS.
So, imagine getting the power, flexibility, stability, and security of a server OS on a desktop? And then, you make it atomic for added security and reliability.
Now we’re talkin’ about something special.
I installed HeliumOS to give it the ol’ test run to see if it was a worthy contender for your desktop. What I found took me by surprise.
My First Impression
My first impression was that HeliumOS’s take on KDE was pretty sweet. It has just the right blend of eye candy and performance. One very nice touch that I noticed is that when using the desktop menu, you get a nice bit of transparency (Figure 1).
Figure 1: The HeliumOS desktop menu is an elegant take on KDE Plasma.
However, if you have an application open and maximized, the menu and panel lose their transparency in favor of reliability (Figure 2).
Figure 2: The transparency vanishes when you have an application open.
Next up is the contents of the desktop menu, which are somewhat spare. You’ll find the Angelfish web browser, which is a QtWebEngine-based browser with DuckDuckGo as the default search engine. This browser is fast, renders well, and strips away a lot of the bloat found in other browsers. You won’t find any signs of AI; there are no extensions, but you can easily create web apps, customize your search engine, and enjoy several ad-blocking lists. That’s about the extent of Angelfish. If that doesn’t suit your needs, you can always open Discover (the KDE Plasma app center) and install whatever browser you want.
Speaking of which…
When you open the app center, you’ll find that most (if not all) apps are of the flatpak type. That suits HeliumOS perfectly well, because the apps are all sandboxed for added security and reliability (which is what this OS is all about).
Installing apps with Discover (even when they are of the Flatpak sort) is as simple as searching for the app you want and clicking Install from Flathub (Figure 3).
Figure 3: Installing an app from Flathub with Discover on HeliumOS.
Another really nifty feature of HeliumOS is that, when you install an application from within Discover, you’ll find an indicator in the menu that something new was installed (Figure 4).
Figure 4: LibreOffice has been installed on HeliumOS.
Support, Support, Support
If you’re looking for a Linux distribution with long-term support, HeliumOS is a great option because it offers 10 years of support, which includes bug fixes, new features, and security updates. Ten years is a long time to rely on an operating system, but HeliumOS wants to make that a reality for you.
But is it worthy of your desktop?
In a word, yes. Normally, I might not be so inclined as to recommend a server-based OS for a desktop, but HeliumOS has changed my mind. That’s not to say HeliumOS is for servers, as it’s not. HeliumOS is only based on a server OS and twisted and tweaked until it is perfectly suited to be your desktop operating system.
After my testing concluded, I realized just how impressed I was with HeliumOS and would happily recommend it as a desktop operating system for anyone from those new to Linux all the way up to seasoned users. HeliumOS is dependable, easy to use, secure, and free. What more do you want from your OS?
And if you want to add gaming into the mix, there’s always Bottles and Steam, both of which can be installed from within Discover.
If you’re interested in trying HeliumOS, download the ISO now and install it on a spare machine or as a virtual machine with VirtualBox.
YOUTUBE.COM/THENEWSTACK
Tech moves fast, don’t miss an episode. Subscribe to our YouTube
channel to stream all our podcasts, interviews, demos, and more.
Jack Wallen is what happens when a Gen Xer mind-melds with present-day snark. Jack is a seeker of truth and a writer of words with a quantum mechanical pencil and a disjointed beat of sound and soul. Although he resides…
